Claude refusal billing resumed 2026-09-24 vs Sume failed runs
Anthropic bills some refusals again from 2026-09-24. How refusal billing works there, and what a failed or cancelled run costs on a Sume Format.

On Claude, a refusal can be billed again: the release note dated 2026-09-24 says billing resumed for the low false-positive refusal categories. A refusal in another category, or with a null category, is not billed. On Sume, the rule is simpler to state: generation that finished before a failure or cancel is billed, and a request rejected at create costs nothing.
Which Claude refusals are billed?
The refusals page ties billing to stop_details.category and to when the refusal happened.
| Case | Billed? |
|---|---|
| Refusal before any output, category bio, frontier_llm or reasoning_extraction | Yes |
| Refusal before any output, any other category or null | No |
| Refusal mid-stream | Input plus the streamed output |
| Rate limits | Refusals still count against them |
What does a failed Sume run cost?
Sume Format billing follows what was generated, not the final status. The docs set these rules:
- A 402 at create (insufficient_credits or organization_wallet_not_provisioned) means nothing ran and nothing is billed.
- A 4xx at create, an idempotent replay and a skipped run cost nothing.
- Generation that finished before a cancel or a failure is billed.
- A run that reaches its generation_spend_cap_usd ends failed, and what it generated is billed.
- The billable figure excludes the agent's own LLM turn, so do not read it as the whole cost of the turn.
How do I see it on a receipt?
The receipt carries usage.billable_amount_usd_micros and usage.generation_spend_cap_usd_micros, so you can compare spend with the cap after the fact. Treat the figure as the generation spend that the docs define, and leave anything outside that definition out of your own totals.
The practical lesson from both vendors is the same. Do not infer cost from the status. A Claude refusal can carry a charge, and a Sume failed run can too. Read the usage fields and keep a cap on each call. The Formats error guide lists the credit and spend codes, and the spend cap section explains the per-run limit.
Sources
Related posts
More in Pricing
- Cost of 1,000 six-second AI video clips by model
A batch of 1,000 six-second clips costs $75 on Grok Imagine Video 1.5 and $750 on Wan 3.0 at 720p. Full Sume table plus how to submit it without queue_full.
- Cost of a 30-second AI video ad: clips, captions, music
A 30-second AI ad costs $0.80 on Grok 720p and up to $6.73 on Kling 3 audio on, once clips, timeline render, captions and music are added. Sume prices.
- Cost per keeper AI video clip: retakes and failed jobs
One usable AI clip costs the clip price divided by your keep rate: at 25 percent, a $0.75 clip costs $3.00 per keeper. Includes a Python function.
- Cost to render a 30-minute video on Sume: a $3.00 ceiling
Sume Timeline reserves $0.10 per output minute and caps a render at 1,800 seconds, so 30 minutes reserves at most $3.00. Length table and the free plan call.
Written by Sume